Documents
Welcome
Introducing
Why Misaki Studio
Compare
Start Here
Learn
Workflows
Export & Ship
AI & MCP
Questions & Answers
Questions from designers
Questions from developers
Questions from React and Next.js developers
UI and MCP questions
Export and boilerplate questions
Accounts and collaboration questions
Reference
Troubleshooting
Examples

UI and MCP questions

Can an agent do the same work as the UI?

Often it can produce the same project result, but it does not reproduce every click or session state. File tabs, panel selection, expansion, zoom, and similar navigation remain UI concerns; some machine inspection and batch operations are MCP-only; several workflows have only partial MCP coverage.

Use the UI-to-MCP capability map.

Can I combine manual editing and agent work?

Yes. Both routes can change the same project model. Work in small scopes: read current state, make one bounded change, verify it, then review the affected file in the editor before beginning the next change.

Follow the combined-work pattern.

How should an agent identify the correct target?

Select the project first and prefer a structured project/file/element reference copied from the editor. Names and paths can be ambiguous after moves or duplicates; reread the target immediately before a write when other people or agents may be editing.

Target files and elements safely.

How do I review what an agent changed?

Review the target's current content, Preview, diagnostics, and generated output as appropriate. Desktop read/write marks remain until explicitly cleared and help locate recent MCP activity, but they are session-only indicators—not authorship, proof, or a durable audit log.

Review agent activity.

Can I undo an agent change?

Usually for supported selected-file content changes. MCP undo and redo use the same design-file history model as the editor and move one history step at a time. They do not roll back file creation/deletion, export destinations, Git, publication, invitations, or other external actions.

Undo agent changes safely.

What proves an agent completed the task?

Evidence appropriate to the result: a read-back for data, Preview or screenshot for appearance, behavior checks for interaction, diagnostics for model/export issues, and the consumer build for shipped code. A successful tool response proves only that the call returned.

Verify agent work.

What if an MCP operation warns or fails halfway through?

Stop and reread the target before retrying. Some operations can report warnings, skipped targets, or partial results; repeated blind writes can compound the problem. Check undo history, Preview, diagnostics, and any external destination affected by the call.

Read MCP limitations.

Does MCP require the desktop app?

The packaged desktop application provides the production MCP server connection. Operations that ask the app to render, preview, capture, or export also require its connected runtime. The browser editor does not independently replace that packaged MCP connection.

Compare browser and desktop.

Can MCP invite users, push project sync, or publish for me?

Not through the current production capability described here. Collaboration invitations, the complete sync/history UI, Git push, package publication, and hosted website publication are separate account or external workflows and must not be inferred from content-authoring tools.

See the exact MCP boundaries.

On this page
UI and MCP questions
Can an agent do the same work as the UI?
Can I combine manual editing and agent work?
How should an agent identify the correct target?
How do I review what an agent changed?
Can I undo an agent change?
What proves an agent completed the task?
What if an MCP operation warns or fails halfway through?
Does MCP require the desktop app?
Can MCP invite users, push project sync, or publish for me?